Aggregation-Based Feature Invention and Relational Concept Classes

Model induction from relational data requires aggregation of values ofattributes of related entities. This paper makes three contributions tothe study of relational learning.(1) It presents a hierarchy ofrelational concepts of increasing complexity, using relational schemacharacteristics such as cardinality, and derives classes of aggregationoperators that are needed to learn these concepts. (2) Expanding onelevel of the hierarchy, it introduces new aggregation operators thatmodel the distribution of the values to be aggregated and (forclassification problems) the differences in these distributions byclass. (3) It demonstrates empirically on a noisy business domain thatmore-complex aggregation methods can increase generalizationperformance. Constructing features using target-dependent aggregationscan transform relational prediction tasks so that well-understoodfeature-vector-based modeling algorithms can be applied successfully.